Discord command replies

Native replies and AI follow-ups

Command answers reply directly to the triggering Discord message, including rich cards and attachments, without pinging its author. Edits keep the answer's message ID when possible; reruns still require the original requester. Explicit cross-channel sends and newly created threads do not reference a message in their parent channel. If the requesting message was deleted, Discord may deliver the answer without a reference; Yetibot does not repeat the command or blindly retry a send.

A plain human reply to a fresh successful standalone !ai, !grok or !zen answer continues with its resolved command/provider/model and safe options. Explicit commands override the branch; plain reply flags/pipes/backticks stay literal. There is no sticky channel mode, inherited actor privilege, bot loop or implicit retry of an error card. Paid routes make new paid requests; automatic free routes stop if the same model is no longer verified free. Edits, deletion, eviction/restart and changed providers expire prior ownership. You can always use an explicit AI command, including for old answers without retained routing.

In ordinary channels/DMs the AI receives that same-channel reply ancestry, including visible V2 answer text, not unrelated channel messages. Actual threads keep their existing full-thread context. The bot needs View Channel and Read Message History. An unreadable or oversized chain fails before inference instead of silently omitting context. See AI conversations for limits and provider/privacy details. Internal model-to-model calls remain stateless and tool-free. Fresh answers establish links; old messages are not rewritten.

Inspect errors and structured results

Show details expands a completed AI reply's frozen tool-call summary inside the same card and becomes Hide details. Hide restores the original answer, files/media and elapsed caption. Long answers may be shortened temporarily while the summary is open. This shows only retained tool names, outcomes and budget counts, not raw arguments/results or private reasoning. It does not rerun tools. Data, tool and error panes replace one another rather than accumulating.

Error details expands retained public diagnostics inside the original reply, without sending another message or rerunning the command. Hide error details restores the original card. Its field/Back/paging controls stay inline. Files and one original elapsed caption are preserved; a long error may be shortened only while details are open. Weather failures can include the service, failure stage/kind and numeric HTTP status. If additional details were not retained, the viewer says so. Raw responses, credentials, headers and private AI internals are not shown. Unexpected Discord command exceptions use a safe summary rather than printing an arbitrary exception message.

Explore data expands inside the original reply and becomes Hide data. It is available for explicitly reviewed public projections: weather, weather forecast, version, changelog, ai tools, AI defaults, provider/model/server/toolset inspection and zen models. AI inspection keeps only reviewed metadata, not credentials, endpoints, raw configs, prompts, schemas or private tool payloads; configured server data does not claim health. The tree initially shows two levels, or all available levels for shallower data. Expand shows one more level everywhere; Collapse shows one fewer. Both reset to the first page. Collapsed objects/arrays show type/count summaries instead of requiring a one-field dropdown. Large JSON previews use Previous and Next at the same depth; pages may continue a document. Controls disable at their bounds. Hide data restores the exact original answer, files/media and elapsed caption. Reopening starts at the two-level overview. Weather snapshot temperatures are labeled Celsius and wind speeds km/h, regardless of the ordinary answer's units. Pipeline data stays unchanged. Raw result maps do not automatically get a public data button.

Snapshots freeze at command completion and are read-only. Public navigation can be used by anyone in the same reply channel/thread; rerunning remains requester-only with confirmation. Edits, cache eviction and process restarts expire old viewers. Inspection never fetches an upstream response or repeats inference/side effects. All panes update only the tracked original reply, with snapshot-token, revision and final ownership checks. Long answers can be shortened temporarily while a pane is open; Hide restores them in full.

Data previews page at1,500 UTF-16 characters, preferably at line boundaries; error navigation retains eight fields/items per page. At most eight data results and eight error records are retained. Snapshots limit nesting, fields/items per container and string sizes; large data may be shortened. Missing or unreviewed fields are not a claim that the upstream response was empty. Adding support requires a reviewed field projection, not blanket map exposure.

Run a command in a thread

Send this directly in a server text channel:

!echo hi | thread
!echo hi | thread | echo result: %s
!weather | thread
!thread echo hello
!thread echo hello | echo result: %s
!thread ai --provider zen Hello
!grok Explain this
!grok Explain this | thread
!thread grok Explain this

The thread is attached to your requesting message. Native | thread passes the piped value/data through unchanged; it does not parse that value as a command. It switches ordinary delivery for later stages and the final answer. Earlier stages have already run, and any output they already sent cannot be moved. Parent working cards are retired when delivery moves; new progress can start in the thread. Explicit cross-channel sends keep their existing semantics.

The prefix form !thread <command> confirms the thread before executing the whole expression, including pipes and embedded commands. Use that form when all work, including initial status/live output, should start in the thread. Grok replies in the current channel/thread by default and creates no thread itself. An intermediate Grok answer is a literal undecorated value for later stages; grok … | thread moves its final output without an earlier independent answer. The legacy --no-thread flag remains accepted and honors explicit thread delivery.

The bot needs permission to create public threads and send messages in them. DMs and nested threads are unsupported. If creation is uncertain, Yetibot may make one bounded read-only lookup to confirm the attached thread; it never retries the creation POST. Without confirmation the thread step fails, no following pipe stages run, and the prefix form does not execute its nested expression. There is no silent parent-channel fallback. Check permissions and send a fresh message. Direct-only admin controls still reject pipes, aliases, wrappers and computed expressions.

Edits and requester-confirmed Run again reuse the same confirmed thread. The requesting user, original message and parent-channel settings remain the source of authorization and context. Moving delivery does not grant permissions. Controls that require direct standalone requests, such as !ai default reset, still reject a thread wrapper. Explicit cross-channel commands retain their existing semantics; thread redirects the ordinary reply destination.

Shared elapsed caption

Rich command answers show a small elapsed footer such as 2.4s, including ordinary commands, errors, media and exact Markdown attachments. It measures the request's command evaluation using a monotonic clock, including preparation of the thread; final transport/cleanup after evaluation is not included. Live answers and working cards show elapsed at publication; terminal reconciliation updates live answer timing to the completed evaluator duration.

AI metadata retains model, known tool count and honest cost labels, but lets the shared renderer own elapsed timing rather than displaying it twice. The caption is transport-only: intermediate pipe values and Slack formatting are unchanged.

The combined rich-text budget remains 4,000 UTF-16 characters. Space is reserved for timing; answers that no longer fit have a readable excerpt and an exact answer.md attachment instead of dropping content.